基于AEDNet的双目立体匹配算法  

Algorithm of binocular stereo matching based on AEDNet

作　　者：杨戈 廖雨婷[1,2] YANG Ge;LIAO Yuting(Advanced Institute of Natural Sciences,Beijing Normal University,Zhuhai 519087,Guangdong China;Key Laboratory of Intelligent Multimedia Technology,Beijing Normal University,Zhuhai 519087,Guangdong China;Engineering Lab on Intelligent Perception for Internet of Things(ELIP),Shenzhen Graduate School,Peking University,Shenzhen 518055,Guangdong China)

机构地区：[1]北京师范大学自然科学高等研究院,广东珠海519087 [2]北京师范大学智能多媒体技术重点实验室,广东珠海519087 [3]北京大学深圳研究生院深圳物联网智能感知技术工程实验室,广东深圳518055

出　　处：《华中科技大学学报（自然科学版）》2022年第3期24-28,共5页Journal of Huazhong University of Science and Technology(Natural Science Edition)

基　　金：国家重点研发计划资助项目(2018YFB1308602);广东高校省级重大科研项目(2018KTSCX288,2019KZDXM015,2020ZDZX3058);广东省学科建设专项资金资助项目(2013WYXM0122);智能多媒体技术重点实验室资助项目(201762005)。

摘　　要：提出了一种基于AANet(有效立体匹配自适应聚合网络)改进的立体匹配网络——AEDNet(自适应端到端立体匹配网络),该网络在特征提取模块通过限制卷积核大小以取得较低抽象程度的特征,通过简单卷积简化网络结构.在代价聚合中采用尺度内聚合模块,通过可变形卷积达到自适应的代价聚合,以及在尺度间聚合模块采用传统的跨尺度聚合方式一定程度上弥补缺失的全局信息.本网络在KITTI数据集上验证其性能,与AANet相比AEDNet参数数量下降了25%,实验结果表明所提网络在保证精度的情况下能高效地实现立体匹配.An improved stereo matching algorithm based on AANet(adaptive aggregation network for efficient stereo matching)was proposed—AEDNet(adaptive end to end depth network for stereo matching). In the feature extraction module,features with low abstraction were obtained by the network by limiting the size of convolution kernel. In the cost aggregation,the intra scale aggregation module was used to achieve adaptive cost aggregation through deformable convolution,and the inter scale aggregation module used the traditional cross scale aggregation method to make up for the missing global information to a certain extent. The network performance was verified on KITTI dataset.Compared with AANet,the number of aednet parameters is reduced by 25%.The experimental results show that the proposed algorithm can achieve stereo matching efficiently with high accuracy.

关 键 词：立体匹配 深度学习 双目立体视觉 卷积神经网络 端到端立体匹配网络 代价聚合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]